Melanlysis: A mobile deep learning approach for early detection of skin cancer

Abstract

Early detection of melanocytes can save lives from melanoma. Most individuals can’t be professionally diagnosed since it’s time-consuming, costly, and inconvenient. Smartphonebased early skin cancer diagnosis has emerged as a new approach. The existing computer-aided skin cancer diagnosis methods and mobile deep learning technology have been studied, and it is found that the existing smartphone-based skin cancer detection and identification methods rely on the support of background cloud services. Accuracy, reaction time, and patient data confidentiality are issues. A novel early detection and recognition model of melanoma skin cancer based on mobile deep learning, Melanlysis, is proposed. The model uses the EfficientNetLite-0 deep learning model to have low latency and considers the imbalance of the existing open-source skin image dataset. The proposed classification model is implemented and evaluated. Experimental results show that compared with the existing EfficientNetLite-0, MobileNet V2, and ResNet-50 models, the accuracy of correctly identifying malignant or non-melanoma is over 94%. At the same time, an Android application based on this mobile deep learning model was developed to diagnose potential malignant melanoma. Users can quickly obtain the classification results of melanoma through the application.